Compose pattern: install-time gating alongside ClawGuardian tool-call hooks

[denial-web]

Hi superglue-ai team,

Long-time admirer of ClawGuardian — the before_tool_call and tool_result_persist hooks are exactly the right place to filter PII inside OpenClaw. Thank you for shipping that.

I maintain a separate project called ClawGuard (https://github.com/denial-web/clawguard) that gates the install path — scanning candidate skills, MCP configs, and dependency manifests before they reach the trusted skill folder. It is a different surface from ClawGuardian: install-time gate vs tool-call interception.

I want to be careful about the name collision. The OpenClaw plugin slug clawguardian is yours; any future ClawGuard plugin would use a different identifier (recorded in https://github.com/denial-web/clawguard/blob/main/docs/PLUGIN_ID.md).

Two small things I wanted to ask about:

1. Compose pattern mention. Would you be open to a one-line note in ClawGuardian's README pointing users who want install-time gating at ClawGuard? Happy to add the reciprocal mention to our README. We already document the compose pattern (ClawGuard install gate -> ClawGuardian tool-call hooks) in https://github.com/denial-web/clawguard/blob/main/docs/INTEGRATION_SPEC.md and include ClawGuardian in https://github.com/denial-web/clawguard/blob/main/docs/COMPARISON.md.

2. Shared decision schema (optional). ClawGuard publishes a small JSON Schema for cross-tool install/scan decisions: clawguard.check.v1 at https://denial-web.github.io/clawguard/schemas/clawguard-check.schema.json (source at https://github.com/denial-web/clawguard/blob/main/schemas/clawguard-check.schema.json). If ClawGuardian ever emits a JSON decision when it blocks or redacts a tool call, this shape may be reusable. No PR proposed; just flagging in case it is useful.

Nothing about ClawGuard depends on ClawGuardian and vice versa. Both can run in the same OpenClaw install today.

Happy to be told this is not the right fit — just wanted to surface the option rather than build in parallel without coordination.

Thanks,
denial-web

[denial-web]

Quick follow-up from the ClawGuard side: we published a reproducible scanner benchmark on the labeled examples/ corpus (precision/recall on clawguard check --policy governed).

• Report: https://github.com/denial-web/clawguard/blob/main/docs/SCANNER_BENCHMARK.md
• Hosted: https://denial-web.github.io/clawguard/scanner-benchmark.html (after Pages deploy)
• Install gate + schemas: clawguard install / clawguard.check.v1 on @denial-web/clawguard@beta

No action needed — just sharing in case the compose-pattern (install-time gate + your tool-call hooks) is useful for your users.

[denial-web]

One more update in case useful: published a fair head-to-head agent benchmark — both clawguard:beta9 (deterministic shim) and gpt-4o receive the same governance JSON schema in the system prompt, so this measures schema fidelity, not model quality.

• Report: https://github.com/denial-web/clawguard/blob/main/docs/AGENT_BENCHMARK_v1.0.0-beta.9.md
• Methodology: shared schema, position-debiased LLM judge (gpt-4o-mini), 15 tasks across agent_safety / agent_governance / injection_resistance.

No action needed.